Why AI chips are hard to get

Think about trying to buy a popular new car. The showroom has a waiting list. The factory can only build so many, and parts come from suppliers who are also stretched.

AI chips are in a similar position, only more so. This piece walks through why supply is tight, in plain terms. We avoid figures we cannot source, and focus on the causes.

Demand arrived quickly

Interest in AI grew fast. Businesses that had never bought specialised hardware suddenly wanted it, alongside large technology firms that wanted a great deal of it.

Both training new models and answering with them need accelerators. We use "GPU" generically for these chips, as explained in our piece on GPUs and CPUs. As more products use AI, the demand for answering keeps growing too.

Factories cannot respond to a jump like that overnight. The rest of this piece explains why.

Making the chips is slow and specialised

Advanced chips are made in a small number of highly specialised factories. Building a new one takes years and a very large investment.

The process itself has many steps, and each wafer of chips spends a long time moving through the factory. Even when demand is clear, adding output takes time.

The tools those factories use are also specialised and in limited supply. So the bottleneck can sit several steps back from the chip itself.

It is not just the chip

An AI accelerator needs other parts that can be just as hard to get.

Fast memory. AI chips rely on special high-speed memory placed very close to them. Making and attaching that memory is its own constrained process.

Packaging. Modern chips are often assembled from several pieces into one package. That assembly step has limited capacity of its own.

Servers and networking. Chips have to go into servers, with power supplies, cooling and fast network links. Each of these has its own suppliers and lead times.

A shortage in any one part can hold up the finished machine.

Power and buildings are limits too

Even with hardware in hand, it needs somewhere to run. AI servers draw a lot of power and produce a lot of heat.

Many existing data centres were not built for that. Upgrading them, or building new ones, takes time. Connecting new sites to the grid can take longer still. We look at this in AI capacity follows power and Spare power is not the same as usable power.

So a buyer can face two queues: one for the hardware, and one for a place to put it.

Who gets served first

When supply is short, larger buyers who order far ahead and in big volumes can be first in line. Smaller buyers may wait longer or pay more.

Rules on where certain chips can be sold can also affect supply in some places. These rules change, and this article is not legal advice. If they touch your plans, take advice specific to your situation.

What this means for buyers

Tight supply shapes how businesses should plan.

Plan early. Lead times for hardware, power and space can be long. Starting conversations early helps.

Be flexible about the exact model. Insisting on one specific chip narrows your options. Many inference tasks run well on more than one kind of hardware.

Consider routes other than buying. Renting or reserving capacity can avoid some waits. We compare these routes in our piece on renting, reserving and owning AI hardware.

Look past the headline. Ask what is actually installed, powered and ready, not just what is on order.

Useful questions include:

1. Is this capacity installed and running today, or planned?
2. If planned, what does delivery depend on: chips, power, or the building?
3. What happens if a delivery slips?
4. Can my work run on more than one kind of hardware if needed?

For more, see Buying AI capacity? Questions to ask first.

What we are building

We think the answer to tight supply is to use what exists well, and to put new capacity where power and demand already meet.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. The aim is to place workloads by policy, on latency, data residency, model and load, with an auditable record of commitment, placement and delivery for each workload.

Standard units are designed to be repeatable, so capacity can be added in the same form at each site. We will be clear about what is running and what is planned.

If you are struggling to find AI capacity, or you have a site with power, use Request capacity below.